Dataset preflight.

Review data issues before you start the next training run. Every finding points to a row and a field.

ERRORS FOUND

Source records4train 3 · eval 1

Errors3Findings that block this profile

Warnings1Review against your task policy

Decoded image files1Unique resolved paths

Findings by rule

image_token_mismatch1

missing_image1

split_image_overlap1

split_sample_overlap1

Counts are findings, not distinct failed records. Showing up to six rules.

Your next steps

Review the affected rows

Resolve errors and decide whether warnings violate your split policy.

Re-run and preserve the evidence

Archive the JSON report, input hashes and experiment configuration.

Run a small training smoke test

Template, tokenizer and model behavior still need separate verification.

Review queue

4 findings

Severity	Rule	Location	What to do
error	`image_token_mismatch`	train:1images	Found 0 <image> markers and 1 image paths; align them.
error	`missing_image`	train:2images[0]	Reference an existing regular image file beneath the image root.
warning	`split_image_overlap`	eval:1images	Identical image bytes appear at train:1; review the split policy.
error	`split_sample_overlap`	eval:1record	Same sample as train:3; remove overlap before evaluating.

Reproducibility manifest · input hashes and settings

{
  "tool": "vlm-data-doctor",
  "version": "0.3.1",
  "profile": "image-text-sft-v2",
  "python": "3.12.15",
  "pillow": "12.3.0",
  "settings": {
    "max_pixels": 40000000,
    "group_key": null
  },
  "dataset_sha256": {
    "train": "59927f7c7f75f8275b3179cba5476dcca0d5a56916f19511b4a0d6f56df464d9",
    "eval": "5393e4e402c0a36f54fa0114d78f3eda714ac9712bda9e9aacd12e217f0c3ca4"
  },
  "decoded_image_inventory_sha256": "86b44209ff53766beb14ffa63f41df2929ec2225e4e3580cef52935b98573bec"
}